平易なコーパスを用いない
テキスト平易化のための
単言語パラレルコーパスの構築
•
English Wikipedia: Alfonso Perez
•
Alfonso Perez Munoz, usually referred to as Alfonso, is a
former Spanish footballer, in the striker position.
•
Simple English Wikipedia: Alfonso Perez
•
Alfonso Perez is a former Spanish football player.
読みやすくなるように文を書き換えるタスク
•
応用1:自然言語処理のために入力文の複雑さを減らす
•
応用2:言語学習者など人々の文章読解を助ける
2
統計的機械翻訳の枠組みでのテキスト平易化
•
テキスト平易化を同一言語内の翻訳問題と考える
•
難解なテキストと平易なテキストからなる
パラレルコーパスを用意してトレーニングする
難解なコーパス
平易なコーパス
統計的機械
翻訳モデル
Some computers can
boot up
from flash drives.
start
Some computers can
from flash drives.
テキスト平易化のための大規模な
言語資源は英語でのみ利用可能
英語
•
平易なコーパス:Simple English Wikipedia(100万文)
•
平易なパラレルコーパス:Parallel Wikipedia(50万文対)
Newsela(5.6万文 5段階の難易度)
•
言い換え辞書:PPDB(1億フレーズ対)
•
平易な言い換え辞書:Simple PPDB(450万フレーズ対)
日本語
•
言い換え辞書:PPDB: Japanese(1,500万フレーズ対)
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
Readability: 80.3 Japan is an island country in East Asia.Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
Readability: 80.3 Japan is an island country in East Asia.Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
1. Japan is a stratovolcanic archipelago of 6,852 islands.
2. Archaeological research indicates that Japan was
inhabited as early as the Upper Paleolithic period.
難解なコーパス
1. Japan is an island country in East Asia.
2. The country is divided into 47 prefectures in eight regions.
3. The population of 126 million is the world's tenth largest.
平易なコーパス
① 文のリーダビリティを求めて難解な文と平易な文に分割
Readability: 80.3 Japan is an island country in East Asia.
Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
1. Japan is a stratovolcanic archipelago of 6,852 islands.
2. Archaeological research indicates that Japan was
inhabited as early as the Upper Paleolithic period.
難解なコーパス
1. Japan is an island country in East Asia.
2. The country is divided into 47 prefectures in eight regions.
3. The population of 126 million is the world's tenth largest.
平易なコーパス
1
1 2 …
0.25
2 3 …
0.11 0.22 0.10 0.00 0.09
文間類似度行列
① 文のリーダビリティを求めて難解な文と平易な文に分割
② 難解な文と平易な文の間の文間類似度を計算
Readability: 80.3 Japan is an island country in East Asia.
Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
1. Japan is a stratovolcanic archipelago of 6,852 islands.
2. Archaeological research indicates that Japan was
inhabited as early as the Upper Paleolithic period.
難解なコーパス
1. Japan is an island country in East Asia.
2. The country is divided into 47 prefectures in eight regions.
3. The population of 126 million is the world's tenth largest.
平易なコーパス
In 599, an earthquake destroyed buildings throughout Yamato Province in what is now Nara Prefecture.
In 599, an earthquake destroyed buildings in Yamato Province which is now known as Nara Prefecture.
パラレルコーパス
1
1 2 …
0.25
2 3 …
0.11 0.22 0.10 0.00 0.09
文間類似度行列
① 文のリーダビリティを求めて難解な文と平易な文に分割
② 難解な文と平易な文の間の文間類似度を計算
③ 閾値以上の文対を抽出してパラレルコーパスを構築
Readability: 80.3 Japan is an island country in East Asia.
Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
1. Japan is a stratovolcanic archipelago of 6,852 islands.
2. Archaeological research indicates that Japan was
inhabited as early as the Upper Paleolithic period.
難解なコーパス
1. Japan is an island country in East Asia.
2. The country is divided into 47 prefectures in eight regions.
3. The population of 126 million is the world's tenth largest.
平易なコーパス
In 599, an earthquake destroyed buildings throughout Yamato Province in what is now Nara Prefecture.
In 599, an earthquake destroyed buildings in Yamato Province which is now known as Nara Prefecture.
パラレルコーパス
統計的機械
翻訳モデル
1
1 2 …
0.25
2 3 …
0.11 0.22 0.10 0.00 0.09
文間類似度行列
① 文のリーダビリティを求めて難解な文と平易な文に分割
② 難解な文と平易な文の間の文間類似度を計算
③ 閾値以上の文対を抽出してパラレルコーパスを構築
④ パラレルコーパスを用いて統計的機械翻訳モデルを学習
Readability: 80.3 Japan is an island country in East Asia.
Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.
平易なコーパスを用いない
テキスト平易化のための単言語パラレルコーパスの構築
1. Japan is a stratovolcanic archipelago of 6,852 islands.
2. Archaeological research indicates that Japan was
inhabited as early as the Upper Paleolithic period.
難解なコーパス
1. Japan is an island country in East Asia.
2. The country is divided into 47 prefectures in eight regions.
3. The population of 126 million is the world's tenth largest.
平易なコーパス
In 599, an earthquake destroyed buildings throughout Yamato Province in what is now Nara Prefecture.
In 599, an earthquake destroyed buildings in Yamato Province which is now known as Nara Prefecture.
パラレルコーパス
統計的機械
翻訳モデル
Some computers can start from flash drives.Some computers can boot up from flash drives.
1
1 2 …
0.25
2 3 …
0.11 0.22 0.10 0.00 0.09
文間類似度行列
① 文のリーダビリティを求めて難解な文と平易な文に分割
② 難解な文と平易な文の間の文間類似度を計算
③ 閾値以上の文対を抽出してパラレルコーパスを構築
④ パラレルコーパスを用いて統計的機械翻訳モデルを学習
⑤ モデルを用いて入力文から平易な同義文を生成
Readability: 80.3 Japan is an island country in East Asia.
Readability: 38.0 Japan is a stratovolcanic archipelago of 6,852 islands. Readability: 69.8 The country is divided into 47 prefectures in eight regions.
Readability: 15.0 Archaeological research indicates that Japan was inhabited as early as the Upper Paleolithic period. Readability: 86.7 The population of 126 million is the world's tenth largest.